I keep seeing these reports/youtubers talking about how NVIDIA is outsmarting the US government by selling chips to China illegally.
This of course doesn't make any sense because NVIDIA is arguably the 3rd most state-embedded company (after Palantir and Microsoft).
The more accurate take is that the US government is managing dependence:
- let enough silicon in to keep China tied to CUDA + NVIDIA networking,
- maintain visibility and leverage, and preserve U.S. scale and R&D speed - while denying full frontier capability.
China's alternatives are improving, but the software + interconnect moat remains the bigger hill to climb. That's why - even with curbs Chinese buyers still chase NVIDIA, and why the U.S. prefers a dial, not a switch.
Why the Controllers still let NVIDIA silicon reach China (even under curbs)
1) Lock-in and dependency > self-sufficiency.
- Let some A/H-class supply drip through and you keep Chinese labs tethered to the CUDA/NVIDIA tooling stack.
- If you cut them off cold, you force a crash program to harden domestic alternatives and a non-CUDA ecosystem.
2) Standards capture.
- Training at frontier scale wants a coherent software + interconnect standard. If China's hyperscalers keep buying - even throttled SKUs - they architect data centers around NVIDIA networking, storage, and scheduling. That makes future policy levers (firmware, driver, interop, licensing) more potent.
3) Telemetry and bargaining chips.
- Allowing legal China-compliant SKUs plus inevitable gray flows gives the US government visibility into where compute concentrates (through export filings, customs, carrier logs, and partner audits).
4) Escalation ladder.
- Partial supply provides a dial (tighten/loosen) instead of a binary switch. That maintains room for response to other events (Taiwan, sanctions, cyber) without jumping straight to maximum escalation.
5) Don't trigger retaliation on crown-jewel U.S. firms.
- A total choke risks Chinese retaliation (licensing, data, taxes, procurement bans) against Apple, Tesla, Qualcomm, Boeing, etc. Allowing constrained flows reduces that risk.
6) Grey-market pressure valve.
- Even with curbs, hardware will transit via re-exports (HK, SG, UAE), second-hand channels, dev kits below thresholds, and capacity rental (cloud time). Designing policy that recognizes - and meters - that reality is smarter than pretending it doesn't exist.
If the US government wanted, they'd dramatically slow "illegal" chip exports to China (tighten end-user lists, ultimate consignee scrutiny, sanctions on third-country distributors, firmware locks).
They don't, because the net-benefit of controlled leakage (dependency + visibility + NVIDIA scale + escalation control) beats the cost of full denial (accelerated Chinese substitutes + retaliation + price spikes in global AI supply chains).
Does China have a credible NVIDIA competitor?
- Short version: Partially, for inference and mid-scale training; not at full frontier scale yet. The moat is software + networking, not just Tera Operations per Second.
So far, the Chinese competitors lack:
- CUDA ecosystem gravity (framework optimizations, libraries, kernels honed over a decade).
- System-level performance.
- Developer familiarity and off-the-shelf tool-chains for every S-curve (ASR, vision, recommender, LLMs, RAG, video).
- Full-stack support (drivers/firmware updates at the cadence hyperscalers expect).
Over a 3–5 year horizon, I'd expect a rising domestic share, but CUDA likely stays first choice where available.
The current mechanisms you see:
- China-specific SKUs (e.g., H20/L20/L2 with restricted performance) officially shipped.
- Third-country intermediaries, re-labeling, and "unrestricted market" routing.
- Cloud capacity rented offshore for Chinese teams (compute by the hour, not by the pallet).
- Refurbished/second-hand boards entering via looser ports.
So China likely remains CUDA-dependent enough to slow self-sovereignty; NVIDIA keeps scale.
At the same time, they'll try to accelerate Huawei Ascend/Cambricon (competitors) and their domestic software stack will likely improve faster.
In the US, competition for NVIDIA seems to be somewhat heating up (but remains to be seen) as it becomes more and more clear that AI governance is the future.
- If the goal is fast, global, uniform governance, NVIDIA wins on installed base + tool-chain + fabric.
- If the goal is sovereign independence from one vendor, the Controllers will co-standardize: NVIDIA plus AMD (ROCm) and a hyperscaler cloud path (AWS/Azure) so there's redundant enforcement.